Skip to contentSkip to conversation

net/netvsc: fix race condition in RNDIS command execution

PATCH In review [email protected] 3 participants 6 messages Jan 13 – Mar 28 Network drivers

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 net/netvsc: fix race condition in RNDIS command execution

From: Madhuker Mythri <[email protected]>

When multiple threads issue RNDIS command requests (such as device info queries and link status checks) simultaneously, command failures can occur due to concurrent access to shared resources in the RNDIS execution path.

Add a spinlock to serialize RNDIS command execution and prevent data corruption.

Fixes 4e9c73e96e83 ("net/netvsc: add Hyper-V network device")
---
 drivers/net/netvsc/hn_ethdev.c | 1 +
 drivers/net/netvsc/hn_rndis.c  | 2 ++
 drivers/net/netvsc/hn_var.h    | 1 +
 3 files changed, 4 insertions(+)
drivers/net/netvsc/hn_ethdev.c +1 -0
index 6584819f4f..b525e287fa 100644
@@ -1310,6 +1310,7 @@ eth_hn_dev_init(struct rte_eth_dev *eth_dev)
 	PMD_INIT_FUNC_TRACE();
 
 	rte_spinlock_init(&hv->hotadd_lock);
+	rte_spinlock_init(&hv->cmd_lock);
 	LIST_INIT(&hv->hotadd_list);
 
 	vmbus = container_of(device, struct rte_vmbus_device, device);
drivers/net/netvsc/hn_rndis.c +2 -0
index 7c54eebcef..8a0716df89 100644
@@ -500,8 +500,10 @@ hn_rndis_query(struct hn_data *hv, uint32_t oid,
 	/* Input data immediately follows RNDIS query. */
 	memcpy(req + 1, idata, idlen);
 
+	rte_spinlock_lock(&hv->cmd_lock);
 	error = hn_rndis_execute(hv, rid, req, reqlen,
 				 comp, comp_len, RNDIS_QUERY_CMPLT);
+	rte_spinlock_unlock(&hv->cmd_lock);
 
 	if (error)
 		goto done;
drivers/net/netvsc/hn_var.h +1 -0
index 17c1d5d07b..66ed186c0a 100644
@@ -179,6 +179,7 @@ struct hn_data {
 	struct vmbus_channel *channels[HN_MAX_CHANNELS];
 
 	rte_spinlock_t	hotadd_lock;
+	rte_spinlock_t	cmd_lock;
 	LIST_HEAD(hotadd_list, hv_hotadd_context) hotadd_list;
 };
 
Signature · 1 line
2.50.1 (Apple Git-155)
#1 lore raw

cover net/netvsc: fix race conditions

The previous patch about netvsc race conditions identified a real issue, but the change introduced another lock.

I propose the following change generated by Claude AI. Unfortunately I no longer have free access to Azure or Hyper-V to test this so please do more complete testing.

Stephen Hemminger (2):
net/netvsc: fix RNDIS command concurrency issue net/netvsc: fix link status RNDIS command concurrency issue

 drivers/net/netvsc/hn_ethdev.c |  13 ++-
 drivers/net/netvsc/hn_rndis.c  | 175 ++++++++++++++++++++-------------
 drivers/net/netvsc/hn_rndis.h  |   2 -
 drivers/net/netvsc/hn_var.h    |   4 +
 4 files changed, 114 insertions(+), 80 deletions(-)
Signature · 1 line
2.51.0
#2 lore raw

PATCH 1/2 net/netvsc: fix RNDIS command concurrency issue

When multiple threads issue RNDIS command requests (such as device info queries) simultaneously, command failures can occur due to concurrent access to shared resources in the RNDIS execution path. The RNDIS layer only supports one pending request at a time, and hn_rndis_exec1() returns -EBUSY if a request is already in progress.

Functions like hn_rndis_get_offload(), hn_rndis_conf_offload(), and hn_rndis_get_ptypes() were querying hardware capabilities on each call, which could race with other RNDIS commands from different threads.

Fix this by querying the hardware offload capabilities once during device attach in hn_rndis_attach() and caching the result in the hn_data structure. All subsequent accesses use the cached value, eliminating the race condition.

If the hwcaps query fails during attach, the device initialization fails entirely. This matches the behavior of the Linux netvsc driver (see rndis_netdev_set_hwcaps in rndis_filter.c).

Fixes 4e9c73e96e83 ("net/netvsc: add Hyper-V network device")
---
 drivers/net/netvsc/hn_rndis.c | 110 ++++++++++++++++------------------
 drivers/net/netvsc/hn_var.h   |   4 ++
 2 files changed, 57 insertions(+), 57 deletions(-)
drivers/net/netvsc/hn_rndis.c +53 -57
index 7c54eebcef..ed053ee905 100644
@@ -781,15 +781,9 @@ int hn_rndis_conf_offload(struct hn_data *hv,
 			  uint64_t tx_offloads, uint64_t rx_offloads)
 {
 	struct ndis_offload_params params;
-	struct ndis_offload hwcaps;
+	const struct ndis_offload *hwcaps = &hv->hwcaps;
 	int error;
 
-	error = hn_rndis_query_hwcaps(hv, &hwcaps);
-	if (error) {
-		PMD_DRV_LOG(ERR, "hwcaps query failed: %d", error);
-		return error;
-	}
-
 	/* NOTE: 0 means "no change" */
 	memset(&params, 0, sizeof(params));
 
@@ -803,25 +797,25 @@ int hn_rndis_conf_offload(struct hn_data *hv,
 	}
 
 	if (tx_offloads & RTE_ETH_TX_OFFLOAD_TCP_CKSUM) {
-		if (hwcaps.ndis_csum.ndis_ip4_txcsum & NDIS_TXCSUM_CAP_TCP4)
+		if (hwcaps->ndis_csum.ndis_ip4_txcsum & NDIS_TXCSUM_CAP_TCP4)
 			params.ndis_tcp4csum = NDIS_OFFLOAD_PARAM_TX;
 		else
 			goto unsupported;
 
-		if (hwcaps.ndis_csum.ndis_ip6_txcsum & NDIS_TXCSUM_CAP_TCP6)
+		if (hwcaps->ndis_csum.ndis_ip6_txcsum & NDIS_TXCSUM_CAP_TCP6)
 			params.ndis_tcp6csum = NDIS_OFFLOAD_PARAM_TX;
 		else
 			goto unsupported;
 	}
 
 	if (rx_offloads & RTE_ETH_RX_OFFLOAD_TCP_CKSUM) {
-		if ((hwcaps.ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_TCP4)
+		if ((hwcaps->ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_TCP4)
 		    == NDIS_RXCSUM_CAP_TCP4)
 			params.ndis_tcp4csum |= NDIS_OFFLOAD_PARAM_RX;
 		else
 			goto unsupported;
 
-		if ((hwcaps.ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_TCP6)
+		if ((hwcaps->ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_TCP6)
 		    == NDIS_RXCSUM_CAP_TCP6)
 			params.ndis_tcp6csum |= NDIS_OFFLOAD_PARAM_RX;
 		else
@@ -829,12 +823,12 @@ int hn_rndis_conf_offload(struct hn_data *hv,
 	}
 
 	if (tx_offloads & RTE_ETH_TX_OFFLOAD_UDP_CKSUM) {
-		if (hwcaps.ndis_csum.ndis_ip4_txcsum & NDIS_TXCSUM_CAP_UDP4)
+		if (hwcaps->ndis_csum.ndis_ip4_txcsum & NDIS_TXCSUM_CAP_UDP4)
 			params.ndis_udp4csum = NDIS_OFFLOAD_PARAM_TX;
 		else
 			goto unsupported;
 
-		if ((hwcaps.ndis_csum.ndis_ip6_txcsum & NDIS_TXCSUM_CAP_UDP6)
+		if ((hwcaps->ndis_csum.ndis_ip6_txcsum & NDIS_TXCSUM_CAP_UDP6)
 		    == NDIS_TXCSUM_CAP_UDP6)
 			params.ndis_udp6csum = NDIS_OFFLOAD_PARAM_TX;
 		else
@@ -842,38 +836,38 @@ int hn_rndis_conf_offload(struct hn_data *hv,
 	}
 
 	if (rx_offloads & RTE_ETH_TX_OFFLOAD_UDP_CKSUM) {
-		if (hwcaps.ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_UDP4)
+		if (hwcaps->ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_UDP4)
 			params.ndis_udp4csum |= NDIS_OFFLOAD_PARAM_RX;
 		else
 			goto unsupported;
 
-		if (hwcaps.ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_UDP6)
+		if (hwcaps->ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_UDP6)
 			params.ndis_udp6csum |= NDIS_OFFLOAD_PARAM_RX;
 		else
 			goto unsupported;
 	}
 
 	if (tx_offloads & RTE_ETH_TX_OFFLOAD_IPV4_CKSUM) {
-		if ((hwcaps.ndis_csum.ndis_ip4_txcsum & NDIS_TXCSUM_CAP_IP4)
+		if ((hwcaps->ndis_csum.ndis_ip4_txcsum & NDIS_TXCSUM_CAP_IP4)
 		    == NDIS_TXCSUM_CAP_IP4)
 			params.ndis_ip4csum = NDIS_OFFLOAD_PARAM_TX;
 		else
 			goto unsupported;
 	}
 	if (rx_offloads & RTE_ETH_RX_OFFLOAD_IPV4_CKSUM) {
-		if (hwcaps.ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_IP4)
+		if (hwcaps->ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_IP4)
 			params.ndis_ip4csum |= NDIS_OFFLOAD_PARAM_RX;
 		else
 			goto unsupported;
 	}
 
 	if (tx_offloads & RTE_ETH_TX_OFFLOAD_TCP_TSO) {
-		if (hwcaps.ndis_lsov2.ndis_ip4_encap & NDIS_OFFLOAD_ENCAP_8023)
+		if (hwcaps->ndis_lsov2.ndis_ip4_encap & NDIS_OFFLOAD_ENCAP_8023)
 			params.ndis_lsov2_ip4 = NDIS_OFFLOAD_LSOV2_ON;
 		else
 			goto unsupported;
 
-		if ((hwcaps.ndis_lsov2.ndis_ip6_opts & HN_NDIS_LSOV2_CAP_IP6)
+		if ((hwcaps->ndis_lsov2.ndis_ip6_opts & HN_NDIS_LSOV2_CAP_IP6)
 		    == HN_NDIS_LSOV2_CAP_IP6)
 			params.ndis_lsov2_ip6 = NDIS_OFFLOAD_LSOV2_ON;
 		else
@@ -898,51 +892,42 @@ int hn_rndis_conf_offload(struct hn_data *hv,
 int hn_rndis_get_offload(struct hn_data *hv,
 			 struct rte_eth_dev_info *dev_info)
 {
-	struct ndis_offload hwcaps;
-	int error;
-
-	memset(&hwcaps, 0, sizeof(hwcaps));
-
-	error = hn_rndis_query_hwcaps(hv, &hwcaps);
-	if (error) {
-		PMD_DRV_LOG(ERR, "hwcaps query failed: %d", error);
-		return error;
-	}
+	const struct ndis_offload *hwcaps = &hv->hwcaps;
 
 	dev_info->tx_offload_capa = RTE_ETH_TX_OFFLOAD_MULTI_SEGS |
 				    RTE_ETH_TX_OFFLOAD_VLAN_INSERT;
 
-	if ((hwcaps.ndis_csum.ndis_ip4_txcsum & HN_NDIS_TXCSUM_CAP_IP4)
+	if ((hwcaps->ndis_csum.ndis_ip4_txcsum & HN_NDIS_TXCSUM_CAP_IP4)
 	    == HN_NDIS_TXCSUM_CAP_IP4)
 		dev_info->tx_offload_capa |= RTE_ETH_TX_OFFLOAD_IPV4_CKSUM;
 
-	if ((hwcaps.ndis_csum.ndis_ip4_txcsum & HN_NDIS_TXCSUM_CAP_TCP4)
+	if ((hwcaps->ndis_csum.ndis_ip4_txcsum & HN_NDIS_TXCSUM_CAP_TCP4)
 	    == HN_NDIS_TXCSUM_CAP_TCP4 &&
-	    (hwcaps.ndis_csum.ndis_ip6_txcsum & HN_NDIS_TXCSUM_CAP_TCP6)
+	    (hwcaps->ndis_csum.ndis_ip6_txcsum & HN_NDIS_TXCSUM_CAP_TCP6)
 	    == HN_NDIS_TXCSUM_CAP_TCP6)
 		dev_info->tx_offload_capa |= RTE_ETH_TX_OFFLOAD_TCP_CKSUM;
 
-	if ((hwcaps.ndis_csum.ndis_ip4_txcsum & NDIS_TXCSUM_CAP_UDP4) &&
-	    (hwcaps.ndis_csum.ndis_ip6_txcsum & NDIS_TXCSUM_CAP_UDP6))
+	if ((hwcaps->ndis_csum.ndis_ip4_txcsum & NDIS_TXCSUM_CAP_UDP4) &&
+	    (hwcaps->ndis_csum.ndis_ip6_txcsum & NDIS_TXCSUM_CAP_UDP6))
 		dev_info->tx_offload_capa |= RTE_ETH_TX_OFFLOAD_UDP_CKSUM;
 
-	if ((hwcaps.ndis_lsov2.ndis_ip4_encap & NDIS_OFFLOAD_ENCAP_8023) &&
-	    (hwcaps.ndis_lsov2.ndis_ip6_opts & HN_NDIS_LSOV2_CAP_IP6)
+	if ((hwcaps->ndis_lsov2.ndis_ip4_encap & NDIS_OFFLOAD_ENCAP_8023) &&
+	    (hwcaps->ndis_lsov2.ndis_ip6_opts & HN_NDIS_LSOV2_CAP_IP6)
 	    == HN_NDIS_LSOV2_CAP_IP6)
 		dev_info->tx_offload_capa |= RTE_ETH_TX_OFFLOAD_TCP_TSO;
 
 	dev_info->rx_offload_capa = RTE_ETH_RX_OFFLOAD_VLAN_STRIP |
 				    RTE_ETH_RX_OFFLOAD_RSS_HASH;
 
-	if (hwcaps.ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_IP4)
+	if (hwcaps->ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_IP4)
 		dev_info->rx_offload_capa |= RTE_ETH_RX_OFFLOAD_IPV4_CKSUM;
 
-	if ((hwcaps.ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_TCP4) &&
-	    (hwcaps.ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_TCP6))
+	if ((hwcaps->ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_TCP4) &&
+	    (hwcaps->ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_TCP6))
 		dev_info->rx_offload_capa |= RTE_ETH_RX_OFFLOAD_TCP_CKSUM;
 
-	if ((hwcaps.ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_UDP4) &&
-	    (hwcaps.ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_UDP6))
+	if ((hwcaps->ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_UDP4) &&
+	    (hwcaps->ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_UDP6))
 		dev_info->rx_offload_capa |= RTE_ETH_RX_OFFLOAD_UDP_CKSUM;
 
 	return 0;
@@ -951,29 +936,20 @@ int hn_rndis_get_offload(struct hn_data *hv,
 uint32_t
 hn_rndis_get_ptypes(struct hn_data *hv)
 {
-	struct ndis_offload hwcaps;
+	const struct ndis_offload *hwcaps = &hv->hwcaps;
 	uint32_t ptypes;
-	int error;
-
-	memset(&hwcaps, 0, sizeof(hwcaps));
-
-	error = hn_rndis_query_hwcaps(hv, &hwcaps);
-	if (error) {
-		PMD_DRV_LOG(ERR, "hwcaps query failed: %d", error);
-		return RTE_PTYPE_L2_ETHER;
-	}
 
 	ptypes = RTE_PTYPE_L2_ETHER;
 
-	if (hwcaps.ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_IP4)
+	if (hwcaps->ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_IP4)
 		ptypes |= RTE_PTYPE_L3_IPV4;
 
-	if ((hwcaps.ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_TCP4) ||
-	    (hwcaps.ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_TCP6))
+	if ((hwcaps->ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_TCP4) ||
+	    (hwcaps->ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_TCP6))
 		ptypes |= RTE_PTYPE_L4_TCP;
 
-	if ((hwcaps.ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_UDP4) ||
-	    (hwcaps.ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_UDP6))
+	if ((hwcaps->ndis_csum.ndis_ip4_rxcsum & NDIS_RXCSUM_CAP_UDP4) ||
+	    (hwcaps->ndis_csum.ndis_ip6_rxcsum & NDIS_RXCSUM_CAP_UDP6))
 		ptypes |= RTE_PTYPE_L4_UDP;
 
 	return ptypes;
@@ -1136,8 +1112,28 @@ hn_rndis_get_linkspeed(struct hn_data *hv)
 int
 hn_rndis_attach(struct hn_data *hv)
 {
+	int error;
+
 	/* Initialize RNDIS. */
-	return hn_rndis_init(hv);
+	error = hn_rndis_init(hv);
+	if (error)
+		return error;
+
+	/*
+	 * Query and cache hardware offload capabilities.
+	 * This avoids the need to query hwcaps on each call to
+	 * hn_rndis_get_offload(), hn_rndis_conf_offload(), or
+	 * hn_rndis_get_ptypes(), which can cause failures if
+	 * multiple threads try to query simultaneously.
+	 */
+	memset(&hv->hwcaps, 0, sizeof(hv->hwcaps));
+	error = hn_rndis_query_hwcaps(hv, &hv->hwcaps);
+	if (error) {
+		PMD_DRV_LOG(ERR, "failed to query hwcaps: %d", error);
+		return error;
+	}
+
+	return 0;
 }
 
 void
drivers/net/netvsc/hn_var.h +4 -0
index 17c1d5d07b..ae7f41dc94 100644
@@ -9,6 +9,8 @@
 #include <rte_eal_paging.h>
 #include <ethdev_driver.h>
 
+#include "ndis.h"
+
 /*
  * Tunable ethdev params
  */
@@ -151,6 +153,8 @@ struct hn_data {
 	uint16_t	num_queues;
 	uint64_t	rss_offloads;
 
+	struct ndis_offload	hwcaps;		/* Cached hardware offload caps */
+
 	rte_spinlock_t	chim_lock;
 	struct rte_mem_resource chim_res;	/* UIO resource for Tx */
 	struct rte_bitmap *chim_bmap;		/* Send buffer map */
Signature · 1 line
2.51.0
#3 lore raw

PATCH 2/2 net/netvsc: fix link status RNDIS command concurrency issue

In reply to Stephen Hemminger

The hn_dev_link_update() function issues RNDIS queries for link status and link speed on every call. Since this function can be called from multiple threads concurrently, it triggers the same RNDIS command concurrency issue where only one request can be pending at a time.

Remove the hn_rndis_get_linkstatus() and hn_rndis_get_linkspeed() functions and modify hn_dev_link_update() to use the cached values directly. The link status and speed are already:

1. Queried once during device attach in hn_rndis_attach()
2. Updated via RNDIS status indications (MEDIA_CONNECT/DISCONNECT and LINK_SPEED_CHANGE) handled in hn_rndis_link_status()

This matches the Linux netvsc driver behavior where link status is queried once in rndis_filter_device_add() and then tracked via status indications.

Fixes 4e9c73e96e83 ("net/netvsc: add Hyper-V network device")
---
 drivers/net/netvsc/hn_ethdev.c | 13 ++++---
 drivers/net/netvsc/hn_rndis.c  | 65 ++++++++++++++++++++++++++--------
 drivers/net/netvsc/hn_rndis.h  |  2 --
 3 files changed, 57 insertions(+), 23 deletions(-)
drivers/net/netvsc/hn_ethdev.c +6 -7
index 6584819f4f..c4580d0bb3 100644
@@ -246,16 +246,15 @@ hn_dev_link_update(struct rte_eth_dev *dev,
 {
 	struct hn_data *hv = dev->data->dev_private;
 	struct rte_eth_link link, old;
-	int error;
 
 	old = dev->data->dev_link;
 
-	error = hn_rndis_get_linkstatus(hv);
-	if (error)
-		return error;
-
-	hn_rndis_get_linkspeed(hv);
-
+	/*
+	 * Use cached link_status and link_speed which are set during
+	 * device attach and updated via RNDIS status indications
+	 * (MEDIA_CONNECT/DISCONNECT and LINK_SPEED_CHANGE).
+	 * This avoids RNDIS queries which are not thread-safe.
+	 */
 	link = (struct rte_eth_link) {
 		.link_duplex = RTE_ETH_LINK_FULL_DUPLEX,
 		.link_autoneg = RTE_ETH_LINK_SPEED_FIXED,
drivers/net/netvsc/hn_rndis.c +51 -14

94 lines not on this page · raw message

drivers/net/netvsc/hn_rndis.h +0 -2
index 7f40f6221d..26854f45a4 100644
@@ -11,8 +11,6 @@ int	hn_rndis_attach(struct hn_data *hv);
 void	hn_rndis_detach(struct hn_data *hv);
 int	hn_rndis_get_eaddr(struct hn_data *hv, uint8_t *eaddr);
 int	hn_rndis_get_mtu(struct hn_data *hv, uint32_t *mtu);
-int	hn_rndis_get_linkstatus(struct hn_data *hv);
-int	hn_rndis_get_linkspeed(struct hn_data *hv);
 int	hn_rndis_set_rxfilter(struct hn_data *hv, uint32_t filter);
 void	hn_rndis_rx_ctrl(struct hn_data *hv, const void *data,
 			 int dlen);
Signature · 1 line
2.51.0
#4 lore raw
In reply to [email protected]
Quoted 9 lines

From: Madhuker Mythri <[email protected]>

When multiple threads issue RNDIS command requests (such as device info queries and link status checks) simultaneously, command failures can occur due to concurrent access to shared resources in the RNDIS execution path.

Add a spinlock to serialize RNDIS command execution and prevent data corruption.

This is not the correct approach as you are only protecting RNDIS_QUERY_CMPLT over RNDIS transactions. The call to hn_rndis_execute() can still fail due to other possible pending transactions over RNDIS.

Thanks,
Long

Quoted 51 lines
Fixes 4e9c73e96e83 ("net/netvsc: add Hyper-V network device")
---
 drivers/net/netvsc/hn_ethdev.c | 1 +
 drivers/net/netvsc/hn_rndis.c  | 2 ++
 drivers/net/netvsc/hn_var.h    | 1 +
 3 files changed, 4 insertions(+)
drivers/net/netvsc/hn_ethdev.c +1 -0
index 6584819f4f..b525e287fa 100644
@@ -1310,6 +1310,7 @@ eth_hn_dev_init(struct rte_eth_dev *eth_dev)
 	PMD_INIT_FUNC_TRACE();
 
 	rte_spinlock_init(&hv->hotadd_lock);
+	rte_spinlock_init(&hv->cmd_lock);
 	LIST_INIT(&hv->hotadd_list);
 
 	vmbus = container_of(device, struct rte_vmbus_device, device); diff --

git a/drivers/net/netvsc/hn_rndis.c b/drivers/net/netvsc/hn_rndis.c index 7c54eebcef..8a0716df89 100644

drivers/net/netvsc/hn_rndis.c +2 -0
@@ -500,8 +500,10 @@ hn_rndis_query(struct hn_data *hv, uint32_t oid,
 	/* Input data immediately follows RNDIS query. */
 	memcpy(req + 1, idata, idlen);
 
+	rte_spinlock_lock(&hv->cmd_lock);
 	error = hn_rndis_execute(hv, rid, req, reqlen,
 				 comp, comp_len, RNDIS_QUERY_CMPLT);
+	rte_spinlock_unlock(&hv->cmd_lock);
 
 	if (error)
 		goto done;
drivers/net/netvsc/hn_var.h index renamed from drivers/net/netvsc/hn_var.h +0 -0

17c1d5d07b..66ed186c0a 100644

drivers/net/netvsc/hn_var.h +1 -1
@@ -179,6 +179,7 @@ struct hn_data {
 	struct vmbus_channel *channels[HN_MAX_CHANNELS];
 
 	rte_spinlock_t	hotadd_lock;
+	rte_spinlock_t	cmd_lock;
 	LIST_HEAD(hotadd_list, hv_hotadd_context) hotadd_list;  };
 
--

2.50.1 (Apple Git-155)

#5 lore raw
In reply to [email protected]
Quoted 7 lines · On Tue, 13 Jan 2026 15:48:20 +0530 [email protected] wrote:

From: Madhuker Mythri <[email protected]>

When multiple threads issue RNDIS command requests (such as device info queries and link status checks) simultaneously, command failures can occur due to concurrent access to shared resources in the RNDIS execution path.

Quoted 8 lines

Add a spinlock to serialize RNDIS command execution and prevent data corruption.

Fixes 4e9c73e96e83 ("net/netvsc: add Hyper-V network device")

---

As Long (and AI review) observed this patch needs more work. There are two issues:

1. The introduced spin lock only protects query, it does not protect around cases where a query races with some thing else calling rndis_set. Lock should go inside hn_rndis_execute().

2. The lock is held while spinning, and it could take several seconds to complete, holding spinlock that long burns CPU and any other thread trying to do operation would queue up and block.

Maybe a pthread_mutex (which sleeps) or use trylock in the query path and return EBUSY?